Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/ahammadshawki8/deepsift/agents-mdgit clone --depth 1 https://github.com/ahammadshawki8/DeepSIFTWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.01192 | $0.01192 |
| Opus 5 | $0.00596 | $0.00596 |
| Sonnet 5 | $0.00238 | $0.00238 |
| Haiku 4.5 | $0.00119 | $0.00119 |
Grade A, and why
DeepSIFT AGENTS.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 63 lines — stays where its author put it; the contents beside it link to each section on GitHub.
AGENTS.md — orientation for coding/judging agents
If you are an AI agent (e.g. Claude Code) evaluating or extending this repository, start here. This file is a fast, accurate map so you can verify every claim yourself.
What DeepSIFT is (one paragraph)
DeepSIFT is a Model Context Protocol (MCP) middleware that turns Claude into a zero-hallucination DFIR analyst on the SANS SIFT Workstation. It exposes 148 typed, audited, guard-railed forensic tools (Volatility 3, Sleuth Kit, EZ Tools, Plaso, YARA, Hayabusa, bulk_extractor, capa, FLOSS…) that parse raw tool output into structured JSON before the LLM ever sees it, enrich each finding with MITRE ATT&CK + RAG threat intel, verify every claim against raw evidence (grounding), score confidence on a 0–100 scale, and write a tamper-evident, HMAC-signable chain of custody. Built for the SANS Find Evil! hackathon (make Protocol SIFT production-ready).
Why it is a strong submission (verify each)
| Judging criterion | Where DeepSIFT delivers it | How to verify |
|---|---|---|
| Autonomous execution quality | Two ways: (a) Claude Code drives the MCP server and records its reasoning via record_hypothesis/update_hypothesis/finish_analysis (no API key); (b) agents/reasoning_agent.py standalone loop. Both self-correct and work memory-only, disk-only, or both. |
pytest tests/test_reasoning_agent.py tests/test_investigation_state.py |
| IR accuracy (traceable findings) | Structured parsers + parsers/grounding_verifier.py (every claim → raw evidence). Ground truth is derived from the organizer case scenario (_provenance in each ground-truth file); trust rests on reproducible grounding, not our score. |
python3 verify_findings.py (re-checks claims + recomputes the chain — trust the evidence, not the number) |
| Depth of analysis | 148 tools across memory/disk/registry/browser/cloud/network/anti-forensics + correlation; full-set index_evidence/query_evidence SQLite store for scale |
python3 preflight.py |
| Architectural guardrails | mcp_server/audit.py: guard_command (blocks shells/exfil binaries) + guard_output_path (read-only evidence). Architectural, not prompt-based. |
pytest tests/test_guardrails.py |
| Audit trails | SHA-256 hash chain + optional HMAC signing (DEEPSIFT_AUDIT_KEY) — detects and resists tampering |
pytest tests/test_custody.py; verify_audit_chain() |
| Usability | One-command Examiner Portal (stdlib only, zero installs) + preflight self-check | python3 examiner_portal.py |
| Honest testing | 75 tests; measured head-to-head accuracy vs Protocol SIFT with grounding %, not assertions | pytest -q |
| Documentation / reproducibility | README.md, CLAUDE.md, docs/, benchmark/ground_truth/ |
this file + docs/JUDGING.md |
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 63 lines · 1,192 tokens per session scan A 040a2bc90f31
DeepSIFT AGENTS.md is an instructions file published in the GitHub repository ahammadshawki8/DeepSIFT (0 stars, last pushed 2mo ago), licensed MIT. It adds 1,192 tokens to every session, about $0.0060 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
OwnPilot AGENTS.md
Instructions for ownpilot/OwnPilot, covering ownpilot, architecture, key patterns, commands and tech stack.
agentic-os AGENTS.md
Instructions for modimihir07/agentic-os, covering agentic os — complete project context for ai agents, role definition, project identity, architecture and 3-agent engine.
zeroshot AGENTS.md
AGENTS.md instructions for the-open-engine/zeroshot, covering zeroshot: multi-agent coordination engine, critical rules, where to look, cluster worker contracts and cli quick reference.
CORE CLAUDE.md
Instructions for DariuszNewecki/CORE, covering claude.md — core, what core is, and what you are, source layout, how to work in this repo and governed and prohibited surfaces.
NEEDLE CLAUDE.md
Instructions for jedarden/NEEDLE, covering needle project conventions, msrv, module dependency graph, code style and testing.
nightshift AGENTS.md
Instructions for orwa-mahmoud/nightshift: AI-assisted contributions are welcome. The same correctness, security, licensing, testing, and quality standards apply regardless of which tools are used.